When I wrote a research note entitled "Latency Matters!" in 2002, I was primarily reacting to the tendency of computer system vendors to highlight how much data they could move around rather than how quickly that data could get from point A to point B. This made comparing server designs--one of my main areas of focus at the time--difficult given that the speed, rather than the amount, of data movement within and between various subsystems was often the more important metric. As I wrote:
Latency is the time that elapses between a request for data and its delivery. It is the sum of the delays each component adds in processing a request. Since it applies to every byte or packet that travels through a system, latency is at least as important as bandwidth, a much-quoted spec whose importance is overrated. High bandwidth just means having a wide, smooth road instead of a bumpy country lane. Latency is the difference between driving it in an old pickup or a Formula One racer.
More recently though, I've seen some great examples that highlight just how important small latency differences are in applications that go well beyond single systems or small clusters. Perhaps unsurprisingly, the financial services industry is driving a lot of this low-latency activity given that trading is all about getting as close to instantaneous as possible.
(I leave the economic and policy aspects of high-speed trading to others. I'd note though that the same page of The Wall Street Journal (PDF) that covers Hibernia Atlantic's announcement also discusses the "flash crash" report.)
One example is the prominence of financial firms like Credit Suisse and JP Morgan Chase in the Advanced Message Queueing Protocol (AMQP) working group. AMQP specifically is being driven by the desire for "open standards efforts between companies to automate electronic transactions are often hindered by the need to incorporate proprietary solutions at the messaging layers of such protocol stacks."
It's also extremely high-performance with native RDMA (Remote Direct Memory Access) Infiniband support, which lets it achieve extremely low end-to-end latencies in the microsecond range. This sort of extremely low latency within a data center was historically associated with smaller complexes of systems and simpler protocols. InfiniBand also continues to be widely used for certain types of high performance computing grids.
However, perhaps the most striking example of the importance of latency even across long distances comes by way of Hibernia Atlantic's announcement on September 30 that they are planning to build the lowest latency cable from New York to London to offer high frequency traders 60 millisecond latencies, which will be the fastest link across the Atlantic.
The first phase of Project Express will begin with a new cable from the county of Somerset in the U.K., to Halifax in Canada where it will then connect to Hibernia's current low latency cable from Halifax to New York. In addition, the new system will include branching units for future latency enhancements to the U.S. and Continental Europe. This work is projected to be completed by the summer of 2012.
Nate Anderson at Ars Technica notes that "operators can plan their geographic routes strategically to keep the total cable length a bit shorter than the competition. According to the consultants at Telegeography, breaking 60ms would make Project Express at least 5ms faster than its closest competitor."
The driving factor here is that, as described by Doug Cameron and Jacob Bunge in the Journal, there's "intense competition to harvest profits from often tiny movements in the price of securities and derivatives." This new transatlantic cable offers a window into how this sort of arbitrage is increasingly global, rather than regional, in scope and is limited only by technology and the laws of physics.
The first transatlantic fiber optic cable in a decade for the purpose of shaving 5 milliseconds off transaction times. Latency very much still matters. Perhaps more than ever.